Meta doubles AI spending to $135B despite slowing agent progress

Meta's massive capital commitment signals that the company views AI infrastructure as a non-negotiable prerequisite for future video and advertising dominance, even as immediate returns on 'agentic' software lag. For the streaming industry, this suggests a bifurcated market where only a few hyperscalers can afford the compute necessary to power next-generation recommendation engines and generative video tools. While competitors like OpenAI and Google have set the pace for foundation models, Meta's strategy relies on brute-force infrastructure spend to catch up. Watch for Muse Spark's upcoming performance benchmarks in complex coding and task-routing as the first indicator of whether this $135 billion investment can produce a credible rival to GPT-5 or Gemini.

The ramp-up in spending follows a volatile period for Meta's capital strategy. Per Seeking Alpha, July 2026, the company’s capital expenditure guidance was recently adjusted to a range of $125 billion to $145 billion, nearly doubling the $72.2 billion spent in 2025. This aggressive spending has drawn comparisons to the Amazon Web Services playbook, with analysts at Morgan Stanley suggesting that if Meta eventually rents even a fraction of this capacity to third parties, it could transform a massive cost center into a high-margin revenue stream.

Recent market activity underscores the potential impact of such a shift. Reports from Bloomberg in July 2026 indicated that Meta is exploring a "Meta Compute" division to sell surplus AI infrastructure. This prospect caused specialized AI cloud providers like CoreWeave and Nebius to see share price drops of 14% and 17% respectively, as investors feared Meta could pivot from a major customer to a direct competitor. CoreWeave currently holds a $21 billion commitment from Meta, while Nebius has a contract worth up to $27 billion, per The Motley Fool, July 2026.

Strategically, Meta is also moving away from its signature open-source approach. While the Llama 4 family (Scout and Maverick) launched in April 2025 with open weights, the new Muse Spark model released in April 2026 is a proprietary, closed-weight system. Developed by the Meta Superintelligence Labs—a division formed after the $14.3 billion acquisition of assets from Scale AI—Muse Spark represents a ground-up rebuild of Meta’s AI stack. According to Artificial Analysis, July 2026, Muse Spark currently ranks in the top five globally for vision and reasoning but continues to struggle in coding benchmarks compared to GPT-5.4 and Claude 4.6.